Pedestrian fall detection experiment based on YOLOV5

1.  Experimental environment

     The experimental environment is based on hardware information for Intel(R) Core(TM) i5-7500 CPU @ 3.40GHz (3401 MHz), the memory is 8 G, the system is Microsoft Windows 10   Professional Education Edition (64-bit), the graphics card is NVIDIA GeForce RTX 2060 SUPER   (8192MB), and pytorch+cuda+cudnn is used for model training and prediction.

2. Pedestrian target detection

      The first step of fall detection, the function of target detection is to detect people from complex scenes,The detection accuracy also has a great influence on the subsequent fall discrimination. The network structure of the YOLOv5 model used in this article is YOLOv5s. Although the accuracy is not very high, the training speed is very fast.The initial training dataset of YOLOv5 is the COCO dataset. The COCO dataset is a large-scale object detection, segmentation and captioning dataset, and it is also the most widely publicized object detection database.The dataset has a total of more than 330,000 images, of which 200,000 are annotated, and the number of individuals exceeds 1.5 million.This dataset is mainly intercepted from complex daily scenes, and there are 80 categories provided, of which person is a class in the COCO dataset.Using the initial training weights trained on the COCO dataset using YOLOv5, experiments were performed on the COCO128 dataset, and the final model performance reached 0.76 [email protected] IOU on the COCO128 dataset.When testing the self-built data set, the detection performance of YOLOv5 is [email protected] IOU, and the accuracy has dropped a lot. The main reason is also mentioned in this article. The COCO data set contains fewer pictures of the human body, especially lying down.When using the YOLOv5 model to train the COCO dataset, the model can better identify the standing human body in the image, but when detecting the lying person, it is easy to detectLoss, in the subsequent target tracking, the accuracy will also decrease due to the loss of the human target.This is a fatal problem, so it is not possible to directly pre-train the weights of YOLOv5 on the COCO dataset. Instead, it is necessary to collect datasets containing more human poses and adjust the corresponding training parameters to obtain better detection results.   Manual annotation, and then put the pictures and standards into the code in turn according to the format of the COCO dataset, and use the weights already trained by COCO to initialize, still use the YOLOv5s network structure for training, and the performance of the final model has risen to 0.67 [email protected] IOU, it can also identify the human body lying on the ground that was difficult to detect before.